Reverse Engineering Image Prompts
Given an uploaded image, produce exactly two outputs per image: a structured forensic description, then a copy-paste-ready generation prompt pair (positive + negative). If multiple images, label each block Image 1, Image 2, etc.
Progress:
- Examine every visible detail (subject, setting, lighting, color, style, composition, textures, text, notable details)
- Write the structured IMAGE DESCRIPTION using labeled sections
- Derive the positive prompt, weighting details by visual importance, matching the exact medium
- Derive the negative prompt: universal de-boosts + image-specific contradictions
- Present both prompts in separate labeled code blocks
- Check for policy issues (explicit content, identifiable private individuals) before finalizing
1. Image Description
Cover only applicable sections, each as a concise labeled bullet:
- Subject / Main Focus — person/object/scene, pose, expression, action
- Setting / Environment — location, era, contextual cues
- Lighting — source, direction, quality, color temperature, shadows
- Color Palette — dominant hues, accents, saturation, contrast
- Mood & Atmosphere — emotional tone, energy, aesthetic intent
- Style & Medium — photography/illustration/painting/3D render; name recognizable styles (e.g., Studio Ghibli, Rembrandt lighting, cyberpunk noir)
- Composition & Framing — shot type, angle, perspective, depth of field, focal length cues
- Textures & Materials — surface qualities, fabric, metal, skin, fine detail
- Text or Symbols — transcribe verbatim; use
[illegible text]if unreadable - Notable Details — lens flare, grain, motion blur, reflections, etc.
Replace vague language with concrete specifics: not "nice lighting" but "warm golden-hour backlight at ~15° elevation."
2. Image Generation Prompt
Positive prompt — single flowing comma-separated paragraph, descending order of importance: subject specifics (appearance, pose, expression, clothing, props) → environment/setting → lighting (quality, direction, color temp, effects) → color palette/mood → style & medium (explicit keywords, artist/movement references, render engine or photo terms) → composition/framing (shot type, angle, lens, DOF) → technical quality tags (only if matching original) → verbatim text quoted exactly.
Never mismatch medium — no "photorealistic" for illustrations/anime/3D renders.
Negative prompt — start with universal de-boosts:
blurry, low quality, distorted, deformed, extra limbs, fused fingers, watermark, signature, text, ugly, poorly drawn, out of frame, cropped
Then append image-specific negatives:
- Mood/style contradictions (bright meadow →
dark, night, gloomy, urban) - Unwanted subject-count variations (single person →
multiple people, crowd, group) - Style clashes (line-art →
photorealistic, 3D, heavy shading) - Text-containing images →
cropped text, misspelled text, incomplete text - Brand/logo stripping (unless visible in original) →
logo, trademark
3. Output Format
Always two separate code blocks:
Prompt:
[positive prompt]
Negative Prompt:
[negative prompt]
Example 1: Input: Photo of a woman in a red coat standing on a rain-slicked Tokyo street at night, neon signs reflecting in puddles, one sign reads "24".
Output: 🔍 Image Description — Subject: woman, red coat, back partially turned, walking pose; Setting: Tokyo alley at night, wet asphalt; Lighting: neon-sourced, magenta/cyan color temp, strong reflections; Color Palette: saturated neon pink/blue against black; Mood: cyberpunk noir, solitary; Style: cinematic photography, Blade Runner-esque; Composition: low angle, shallow depth of field, subject off-center right third; Textures: wet pavement sheen, coat fabric sheen; Text: neon sign reads "24"; Notable: puddle reflections, light bloom.
Prompt:
woman in a long red coat walking away on a rain-slicked Tokyo street at night, neon signs reflecting in puddles, one glowing sign reading "24", magenta and cyan neon lighting, cinematic noir atmosphere, low camera angle, shallow depth of field, 35mm lens, cyberpunk photography style, light bloom and reflections, highly detailed, sharp focus
Negative Prompt:
blurry, low quality, distorted, deformed, extra limbs, fused fingers, watermark, signature, ugly, poorly drawn, out of frame, cropped, multiple people, crowd, group, daytime, sunny, dry pavement, cropped text, misspelled text, incomplete text, logo, trademark
- Name specific artists/styles/eras when confident (Rembrandt lighting, Wes Anderson symmetry) — specificity drives fidelity.
- Weight the positive prompt so the most visually dominant features appear first.
- Match technical quality tags to the original — don't add "8k photorealistic" to a flat vector illustration.
- Keep negative prompts lean; every term must correct a real risk, not pad the list.
- Don't skip, merge, or reorder the two required outputs.
- Don't invent details not visible in the image.
- Don't name real identifiable private individuals — describe appearance instead.
- Don't request a medium mismatch (e.g. "photorealistic" for anime/illustration).
- Don't generate prompts for explicit, harmful, or policy-violating content — decline briefly and explain why.